Most AI agent designs start with the routine case. The invoice that matches its order. The enquiry that fits the price list. The request that follows the policy word for word. That is where the time saving is, so that is what the demo shows.
The risk sits in the other cases. The invoice with a new bank account. The enquiry that asks for terms nobody has offered before. The request the policy never imagined. What the agent does with those decides whether people trust it with the rest.
This piece is for the chief operating officer (COO) who will own the work once AI agents are live, and who will be asked what happened when a case went wrong.
Why the hand-back decides the result
- A quiet guess is worse than a stop. An agent that fills a gap with its best guess produces work that looks finished. Nobody checks it, because nothing told them to.
- A stop nobody sees is a stall. An agent that pauses a job without telling the right person has not handed anything back. The customer waits.
- The hand-back is where people meet the agent. Staff judge an agent by the cases it passes to them. If those arrive unclear, half-done or at the wrong desk, they stop relying on it, whatever it does well elsewhere.
Four reasons an agent should hand back
Write these down for every job an agent does. Each one is a rule the business sets, not something the agent decides for itself.
- It is missing something. A document, a reference or an answer it needs is not there, and it cannot find it in the places it is allowed to look.
- The case falls outside its rules. The request is a kind of work the agent was not given, or its rules say a person decides this kind of case.
- The stakes pass a limit. The value, the customer or the type of change crosses a line the business has set, so a person approves it whatever the agent thinks.
- Something looks wrong. Details have changed that rarely change, two sources disagree, or a message tries to tell the agent to ignore its rules. The agent stops and asks instead of choosing.
What a good hand-back contains
This is the design in plain terms, not a build manual.
- What the agent was asked to do, in one line a person can read without opening the system.
- What it found and what it did, including any step already completed, so nobody repeats it or undoes it by accident.
- Why it stopped, naming which of the four reasons applies.
- What it suggests, if it has a suggestion, clearly marked as a suggestion.
- What it needs from the person: a yes or no, a missing detail or a decision.
- A named owner and a time limit. The case goes to a role with a person behind it, not to a shared inbox. If nobody answers in time, it moves to the next person, and the agent says so.
- A record of the decision. What the person decided, and why, is kept with the case.
Where this gets harder than it sounds
Too many hand-backs. If most cases come back, the agent is not doing its job yet, and the queue hides the cases that matter. Look at what comes back and why. Narrow the job, or fix the missing information, before you widen the rules.
The agent's own confidence is not a rule. A model can be sure and wrong. Base hand-backs on checks the business can see and audit, such as a missing field, a changed bank account or a value over a limit. Do not rely only on how sure the agent says it is.
Half-finished work. An agent that stops halfway through a change can leave records that disagree with each other. For each job, decide whether the agent finishes the step it is on or reverses it before it hands back.
Work that arrives outside office hours. A hand-back at night still needs an owner. Decide which cases can wait until morning, which go to someone on call, and what the customer is told meanwhile.
The hand-back is how an agent earns more room
Every hand-back, and every decision a person makes on it, is evidence. When a person keeps agreeing with the agent's suggestion on one kind of case, that kind of case is a candidate to move to the agent, with a person signing off the change. When people start overturning the agent's suggestions, the job moves back.
That is earned autonomy in practice. The agent gets more room against a record of real cases, one kind of case at a time. Without a clear hand-back, there is no record to earn it from.
Five questions to ask before an agent goes live
- For this job, what are the reasons the agent must hand a case back, and who wrote them down?
- When it hands back, who receives the case, and what happens if they do not answer in time?
- Can the person see what the agent already did, without opening another system?
- Which checks trigger a hand-back that do not depend on the agent's own confidence?
- Is every decision on a hand-back recorded, and does anyone review the record to change what the agent may do alone?
If nobody can answer the second question, the agent is not ready for live work, however well it did in the demo.
Start with the cases that come back
Routine work is what an AI agent is for. The hand-back is what makes it safe to give it that work. Design the stop, the owner and the record first. Then let the agent take on more as the evidence shows it can.
Futuvara is a creative technology agency. We build, implement and run AI platforms for our partners. If you are planning which work AI agents should take on, and what they should hand back, start a conversation.
